Planroom

Construction-document analysis tested against real bid packages

Services: Document AI ; RAG Knowledge Systems ; AI Agents ; AI Workflow Automation
Delivery: Document processing, structured extraction, and project-scoped retrieval. Portfolio alias.

These figures describe evaluation scope, not perfect extraction or production throughput.

Bid information can be scattered across an invitation email, a drawing title block, a specification, and an addendum. The team needs more than a summary. It needs the deadline, scope, exclusions, compliance obligations, and the source behind each item.

Large documents made the problem harder. A field could be extracted somewhere in the output but fail to appear in the field the user needed. The March review found examples where a bid date existed in an entries list but was absent from the headline bid-due field.

That is a product correctness problem, even when the underlying text was read successfully.

The platform organizes projects, documents, permissions, and team assignments. Processing services extract content, produce chunks and summaries, and build retrieval indexes. Evidence records preserve the connection to source material.

A project-scoped question endpoint passes tenant and project context to a Python retrieval worker. Its LangGraph flow separates tool calls, retrieval review, and further reasoning before returning an answer.

Structured extraction covers scope, quantities, schedule, compliance, cost, risk, and related project information. The ten-pack review compared outputs with the original documents rather than accepting fluent text as sufficient evidence.

The review covered PDFs, office files, emails, screenshots, and plan sets. It reported strong project-identity capture, but also material gaps in deadlines, insurance, bonding, and risk outputs.

Those findings explain the need for field-level checks and source review. A result can look detailed while missing the one deadline that determines whether a contractor can submit a bid.

The case study does not claim that the historical gaps have all been eliminated. Current acceptance results are needed before attaching an accuracy percentage.

The delivered system provides a project workspace, document-processing pipeline, structured outputs, retrieval, and evidence inspection. The available review establishes that it was exercised against fifteen real construction packages, including the large package described above.

The engineering record supports a tested document workflow. It does not support an error-free estimating claim.

React/Next.js; Node.js, Express, TypeScript; PostgreSQL and S3; OCR/parser workers; Pinecone and embedding services; Python, LangGraph, OpenAI Responses API; pipeline checkpoints, error records, and telemetry.

From bid documents to structured answers with source evidence.
Planroom combines construction-document processing and project-scoped AI retrieval. Its extraction review covered fifteen real packages, including a sixty-two-document package with a 1,543-page specification.

We built a construction-document system and evaluated it against actual bid packages. The difficult part was carrying the right evidence into the right output field, so the work includes extraction testing and source inspection as well as AI answers.